From: thule@thule.invalid   
      
   On 10/27/25 12:56 AM, an0n wrote:   
   > Interesting article shared on Y-Combinator > https://arxiv.org/   
   > abs/2510.18212 trying to quantify how close we are to AGI by starting   
   > with an empirical analysis of basic human intelligence.   
   >   
   > Tl;Dr > The paper attributes overall AGI score of 27% for GPT-4 and 57%   
   > for GPT-5, showing both the rapid progress and the substantial gaps   
   > remaining to reach human-level cognition. The most significant   
   > bottleneck identified is Long-Term Memory Storage (MS); This fundamental   
   > "amnesia" means models must rely on "capability contortions", such as   
   > using massive context windows (Working Memory) to compensate for their   
   > inability to learn and store new information persistently, often the   
   > root cause of hallucinations.   
   >   
   > Soooooo .... not long now boys !! "just" need to solve for these issues   
   > and we are off to the races.   
      
   i'm not one of those ai-hating redditors; i host a few small models on   
   my home network and use them regularly.  i'm unqualified but i have read   
   a few computational neuroscience textbooks and i'm convinced llms cannot   
   become agi.  a fundamental limitation is their lack of world models, and   
   some models now work around this with chain of thought, but it's not a   
   true solution.  while general intelligence is defined as whatever humans   
   can do, i doubt there's such thing as truly "general" - i.e. opposite of   
   narrow - intelligence; indeed, the brain is a collection of specialised   
   regions bootstrapped together, with the cortex and its canonical circuit   
   being the only sort of general-purpose structure i'm aware of.   
      
   --- SoupGate-Win32 v1.05   
    * Origin: you cannot sedate... all the things you hate (1:229/2)